Expandable and Differentiable Dual Memories with Orthogonal Regularization for Exemplar-free Continual Learning